Research on Harmonic Prediction Model Based on Data-Driven Method

Aiming at the demand of harmonic data quantification and in-depth analysis in power systems, this paper proposes a harmonic data prediction method based on VMD-DeepAR-SOFTS combined model. Firstly, the complex nonlinear and non-stationary harmonic signal was decomposed into multiple Intrinsic mode Functions (IMFs) with different frequency characteristics by using Variational Mode Decomposition (VMD), which effectively improved the separability of the signal and reduced the noise interference. Then, the DeepAR model is used to predict the time series of each IMF component, and the sequential feature selection technology SOFTS based on window optimization is combined to further improve the efficiency of feature extraction and the accuracy of prediction. Experimental results show that the VMD-DeepAR-SOFTS combined model achieves 0.0128, 0.9099 and 0.015523 in MAE, R² and RMSE, respectively, which is significantly better than traditional machine learning models such as LightGBM, XGBoost, CatBoost and SVR. In addition, through the verification of ten groups of independent data sets randomly derived from the system PS1000, the model shows a high degree of consistency and stability, which verifies its excellent generalization ability and robustness. Compared with the single DeepAR or SOFTS model, the combined model has a significant improvement in prediction accuracy and real-time performance. The proposed method not only improves the accuracy of harmonic prediction, reduces the dependence on model parameter tuning, reduces the complexity and cost in practical applications, but also demonstrates its broad application prospects in complex power system environments. Future research will further optimize the model structure, explore more advanced time series decomposition and feature selection techniques to improve the performance of the model, and verify its applicability and effectiveness in more actual power system scenarios.